A New Scene Transfer Method for Flame Burning Images Based on Generative Adversarial Networks

The training of video fire detection models based on deep learning relies on a large number of positive and negative samples, namely fire video and scenario video with other disturbances similar to fire. In some special occasions, the fire video sampled in this scene is insufficient due to the prohibition of ignition. In this thesis, the migration of flames recorded in other similar scenarios to the specific scenarios is proposed to increase the video of fires in restricted situations. Considering the reservation and transmission of information, the generative adversarial networks is adopted and the defects of the existing methods of disentangling image information into latent spaces are overcome. This model can add flames to a given scene video and complete the fusion of style information while maintaining high quality background details. Experimental results show that this method can ensure the diversity of fire in the scene, and produce perceptually realistic results.